File size: 20,370 Bytes
9aa5185 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 | ---
sidebar_position: 5
title: "Environments, Benchmarks & Data Generation"
description: "Building RL training environments, running evaluation benchmarks, and generating SFT data with the Hermes-Agent Atropos integration"
---
# Environments, Benchmarks & Data Generation
Hermes Agent includes a full environment framework that connects its tool-calling capabilities to the [Atropos](https://github.com/NousResearch/atropos) RL training framework. This enables three workflows:
1. **RL Training** β Train language models on multi-turn agentic tasks with GRPO
2. **Benchmarks** β Evaluate models on standardised agentic benchmarks
3. **Data Generation** β Generate SFT training data from agent rollouts
All three share the same core: an **environment** class that defines tasks, runs an agent loop, and scores the output.
:::info Repo environments vs RL training tools
The Python environment framework documented here lives under the repo's `environments/` directory and is the implementation-level API for Hermes/Atropos integration. This is separate from the user-facing `rl_*` tools, which operate as an orchestration surface for remote RL training workflows.
:::
:::tip Quick Links
- **Want to run benchmarks?** Jump to [Available Benchmarks](#available-benchmarks)
- **Want to train with RL?** See [RL Training Tools](/user-guide/features/rl-training) for the agent-driven interface, or [Running Environments](#running-environments) for manual execution
- **Want to create a new environment?** See [Creating Environments](#creating-environments)
:::
## Architecture
The environment system is built on a three-layer inheritance chain:
```mermaid
classDiagram
class BaseEnv {
Server management
Worker scheduling
Wandb logging
CLI: serve / process / evaluate
}
class HermesAgentBaseEnv {
Terminal backend configuration
Tool resolution
Agent loop engine
ToolContext access
}
class TerminalTestEnv {
Stack testing
}
class HermesSweEnv {
SWE training
}
class TerminalBench2EvalEnv {
Benchmark evaluation
}
class TBLiteEvalEnv {
Fast benchmark
}
class YCBenchEvalEnv {
Long-horizon benchmark
}
BaseEnv <|-- HermesAgentBaseEnv
HermesAgentBaseEnv <|-- TerminalTestEnv
HermesAgentBaseEnv <|-- HermesSweEnv
HermesAgentBaseEnv <|-- TerminalBench2EvalEnv
TerminalBench2EvalEnv <|-- TBLiteEvalEnv
TerminalBench2EvalEnv <|-- YCBenchEvalEnv
```
### BaseEnv (Atropos)
The foundation from `atroposlib`. Provides:
- **Server management** β connects to OpenAI-compatible APIs (VLLM, SGLang, OpenRouter)
- **Worker scheduling** β parallel rollout coordination
- **Wandb integration** β metrics logging and rollout visualisation
- **CLI interface** β three subcommands: `serve`, `process`, `evaluate`
- **Eval logging** β `evaluate_log()` saves results to JSON + JSONL
### HermesAgentBaseEnv
The hermes-agent layer (`environments/hermes_base_env.py`). Adds:
- **Terminal backend configuration** β sets `TERMINAL_ENV` for sandboxed execution (local, Docker, Modal, Daytona, SSH, Singularity)
- **Tool resolution** β `_resolve_tools_for_group()` calls hermes-agent's `get_tool_definitions()` to get the right tool schemas based on enabled/disabled toolsets
- **Agent loop integration** β `collect_trajectory()` runs `HermesAgentLoop` and scores the result
- **Two-phase operation** β Phase 1 (OpenAI server) for eval/SFT, Phase 2 (VLLM ManagedServer) for full RL with logprobs
- **Async safety patches** β monkey-patches Modal backend to work inside Atropos's event loop
### Concrete Environments
Your environment inherits from `HermesAgentBaseEnv` and implements five methods:
| Method | Purpose |
|--------|---------|
| `setup()` | Load dataset, initialise state |
| `get_next_item()` | Return the next item for rollout |
| `format_prompt(item)` | Convert an item into the user message |
| `compute_reward(item, result, ctx)` | Score the rollout (0.0β1.0) |
| `evaluate()` | Periodic evaluation logic |
## Core Components
### Agent Loop
`HermesAgentLoop` (`environments/agent_loop.py`) is the reusable multi-turn agent engine. It runs the same tool-calling pattern as hermes-agent's main loop:
1. Send messages + tool schemas to the API via `server.chat_completion()`
2. If the response contains `tool_calls`, dispatch each via `handle_function_call()`
3. Append tool results to the conversation, go back to step 1
4. If no `tool_calls`, the agent is done
Tool calls execute in a thread pool (`ThreadPoolExecutor(128)`) so that async backends (Modal, Docker) don't deadlock inside Atropos's event loop.
Returns an `AgentResult`:
```python
@dataclass
class AgentResult:
messages: List[Dict[str, Any]] # Full conversation history
turns_used: int # Number of LLM calls made
finished_naturally: bool # True if model stopped on its own
reasoning_per_turn: List[Optional[str]] # Extracted reasoning content
tool_errors: List[ToolError] # Errors encountered during tool dispatch
managed_state: Optional[Dict] # VLLM ManagedServer state (Phase 2)
```
### Tool Context
`ToolContext` (`environments/tool_context.py`) gives reward functions direct access to the **same sandbox** the model used during its rollout. The `task_id` scoping means all state (files, processes, browser tabs) is preserved.
```python
async def compute_reward(self, item, result, ctx: ToolContext):
# Run tests in the model's terminal sandbox
test = ctx.terminal("pytest -v")
if test["exit_code"] == 0:
return 1.0
# Check if a file was created
content = ctx.read_file("/workspace/solution.py")
if content.get("content"):
return 0.5
# Download files for local verification
ctx.download_file("/remote/output.bin", "/local/output.bin")
return 0.0
```
Available methods:
| Category | Methods |
|----------|---------|
| **Terminal** | `terminal(command, timeout)` |
| **Files** | `read_file(path)`, `write_file(path, content)`, `search(query, path)` |
| **Transfers** | `upload_file()`, `upload_dir()`, `download_file()`, `download_dir()` |
| **Web** | `web_search(query)`, `web_extract(urls)` |
| **Browser** | `browser_navigate(url)`, `browser_snapshot()` |
| **Generic** | `call_tool(name, args)` β escape hatch for any hermes-agent tool |
| **Cleanup** | `cleanup()` β release all resources |
### Tool Call Parsers
For **Phase 2** (VLLM ManagedServer), the server returns raw text without structured tool calls. Client-side parsers in `environments/tool_call_parsers/` extract `tool_calls` from raw output:
```python
from environments.tool_call_parsers import get_parser
parser = get_parser("hermes") # or "mistral", "llama3_json", "qwen", "deepseek_v3", etc.
content, tool_calls = parser.parse(raw_model_output)
```
Available parsers: `hermes`, `mistral`, `llama3_json`, `qwen`, `qwen3_coder`, `deepseek_v3`, `deepseek_v3_1`, `kimi_k2`, `longcat`, `glm45`, `glm47`.
In Phase 1 (OpenAI server type), parsers are not needed β the server handles tool call parsing natively.
## Available Benchmarks
### TerminalBench2
**89 challenging terminal tasks** with per-task Docker sandbox environments.
| | |
|---|---|
| **What it tests** | Single-task coding/sysadmin ability |
| **Scoring** | Binary pass/fail (test suite verification) |
| **Sandbox** | Modal cloud sandboxes (per-task Docker images) |
| **Tools** | `terminal` + `file` |
| **Tasks** | 89 tasks across multiple categories |
| **Cost** | ~$50β200 for full eval (parallel execution) |
| **Time** | ~2β4 hours |
```bash
python environments/benchmarks/terminalbench_2/terminalbench2_env.py evaluate \
--config environments/benchmarks/terminalbench_2/default.yaml
# Run specific tasks
python environments/benchmarks/terminalbench_2/terminalbench2_env.py evaluate \
--config environments/benchmarks/terminalbench_2/default.yaml \
--env.task_filter fix-git,git-multibranch
```
Dataset: [NousResearch/terminal-bench-2](https://huggingface.co/datasets/NousResearch/terminal-bench-2) on HuggingFace.
### TBLite (OpenThoughts Terminal Bench Lite)
**100 difficulty-calibrated tasks** β a faster proxy for TerminalBench2.
| | |
|---|---|
| **What it tests** | Same as TB2 (coding/sysadmin), calibrated difficulty tiers |
| **Scoring** | Binary pass/fail |
| **Sandbox** | Modal cloud sandboxes |
| **Tools** | `terminal` + `file` |
| **Tasks** | 100 tasks: Easy (40), Medium (26), Hard (26), Extreme (8) |
| **Correlation** | r=0.911 with full TB2 |
| **Speed** | 2.6β8Γ faster than TB2 |
```bash
python environments/benchmarks/tblite/tblite_env.py evaluate \
--config environments/benchmarks/tblite/default.yaml
```
TBLite is a thin subclass of TerminalBench2 β only the dataset and timeouts differ. Created by the OpenThoughts Agent team (Snorkel AI + Bespoke Labs). Dataset: [NousResearch/openthoughts-tblite](https://huggingface.co/datasets/NousResearch/openthoughts-tblite).
### YC-Bench
**Long-horizon strategic benchmark** β the agent plays CEO of an AI startup.
| | |
|---|---|
| **What it tests** | Multi-turn strategic coherence over hundreds of turns |
| **Scoring** | Composite: `0.5 Γ survival + 0.5 Γ normalised_funds` |
| **Sandbox** | Local terminal (no Modal needed) |
| **Tools** | `terminal` only |
| **Runs** | 9 default (3 presets Γ 3 seeds), sequential |
| **Cost** | ~$50β200 for full eval |
| **Time** | ~3β6 hours |
```bash
# Install yc-bench (optional dependency)
pip install "hermes-agent[yc-bench]"
# Run evaluation
bash environments/benchmarks/yc_bench/run_eval.sh
# Or directly
python environments/benchmarks/yc_bench/yc_bench_env.py evaluate \
--config environments/benchmarks/yc_bench/default.yaml
# Quick single-preset test
python environments/benchmarks/yc_bench/yc_bench_env.py evaluate \
--config environments/benchmarks/yc_bench/default.yaml \
--env.presets '["fast_test"]' --env.seeds '[1]'
```
YC-Bench uses [collinear-ai/yc-bench](https://github.com/collinear-ai/yc-bench) β a deterministic simulation with 4 skill domains (research, inference, data_environment, training), prestige system, employee management, and financial pressure. Unlike TB2's per-task binary scoring, YC-Bench measures whether an agent can maintain coherent strategy over hundreds of compounding decisions.
## Training Environments
### TerminalTestEnv
A minimal self-contained environment with inline tasks (no external dataset). Used for **validating the full stack** end-to-end. Each task asks the model to create a file at a known path; the verifier checks the content.
```bash
# Process mode (saves rollouts to JSONL, no training server needed)
python environments/terminal_test_env/terminal_test_env.py process \
--env.data_path_to_save_groups terminal_test_output.jsonl
# Serve mode (connects to Atropos API for RL training)
python environments/terminal_test_env/terminal_test_env.py serve
```
### HermesSweEnv
SWE-bench style training environment. The model gets a coding task, uses terminal + file + web tools to solve it, and the reward function runs tests in the same Modal sandbox.
```bash
python environments/hermes_swe_env/hermes_swe_env.py serve \
--openai.model_name YourModel \
--env.dataset_name bigcode/humanevalpack \
--env.terminal_backend modal
```
## Running Environments
Every environment is a standalone Python script with three CLI subcommands:
### `evaluate` β Run a benchmark
For eval-only environments (benchmarks). Runs all items, computes metrics, logs to wandb.
```bash
python environments/benchmarks/tblite/tblite_env.py evaluate \
--config environments/benchmarks/tblite/default.yaml \
--openai.model_name anthropic/claude-sonnet-4.6
```
No training server or `run-api` needed. The environment handles everything.
### `process` β Generate SFT data
Runs rollouts and saves scored trajectories to JSONL. Useful for generating training data without a full RL loop.
```bash
python environments/terminal_test_env/terminal_test_env.py process \
--env.data_path_to_save_groups output.jsonl \
--openai.model_name anthropic/claude-sonnet-4.6
```
Output format: each line is a scored trajectory with the full conversation history, reward, and metadata.
### `serve` β Connect to Atropos for RL training
Connects the environment to a running Atropos API server (`run-api`). Used during live RL training.
```bash
# Terminal 1: Start the Atropos API
run-api
# Terminal 2: Start the environment
python environments/hermes_swe_env/hermes_swe_env.py serve \
--openai.model_name YourModel
```
The environment receives items from Atropos, runs agent rollouts, computes rewards, and sends scored trajectories back for training.
## Two-Phase Operation
### Phase 1: OpenAI Server (Eval / SFT)
Uses `server.chat_completion()` with `tools=` parameter. The server (VLLM, SGLang, OpenRouter, OpenAI) handles tool call parsing natively. Returns `ChatCompletion` objects with structured `tool_calls`.
- **Use for**: evaluation, SFT data generation, benchmarks, testing
- **Placeholder tokens** are created for the Atropos pipeline (since real token IDs aren't available from the OpenAI API)
### Phase 2: VLLM ManagedServer (Full RL)
Uses ManagedServer for exact token IDs + logprobs via `/generate`. A client-side [tool call parser](#tool-call-parsers) reconstructs structured `tool_calls` from raw output.
- **Use for**: full RL training with GRPO/PPO
- **Real tokens**, masks, and logprobs flow through the pipeline
- Set `tool_call_parser` in config to match your model's format (e.g., `"hermes"`, `"qwen"`, `"mistral"`)
## Creating Environments
### Training Environment
```python
from environments.hermes_base_env import HermesAgentBaseEnv, HermesAgentEnvConfig
from atroposlib.envs.server_handling.server_manager import APIServerConfig
class MyEnvConfig(HermesAgentEnvConfig):
my_custom_field: str = "default_value"
class MyEnv(HermesAgentBaseEnv):
name = "my-env"
env_config_cls = MyEnvConfig
@classmethod
def config_init(cls):
env_config = MyEnvConfig(
enabled_toolsets=["terminal", "file"],
terminal_backend="modal",
max_agent_turns=30,
)
server_configs = [APIServerConfig(
base_url="https://openrouter.ai/api/v1",
model_name="anthropic/claude-sonnet-4.6",
server_type="openai",
)]
return env_config, server_configs
async def setup(self):
from datasets import load_dataset
self.dataset = list(load_dataset("my-dataset", split="train"))
self.iter = 0
async def get_next_item(self):
item = self.dataset[self.iter % len(self.dataset)]
self.iter += 1
return item
def format_prompt(self, item):
return item["instruction"]
async def compute_reward(self, item, result, ctx):
# ctx gives full tool access to the rollout's sandbox
test = ctx.terminal("pytest -v")
return 1.0 if test["exit_code"] == 0 else 0.0
async def evaluate(self, *args, **kwargs):
# Periodic evaluation during training
pass
if __name__ == "__main__":
MyEnv.cli()
```
### Eval-Only Benchmark
For benchmarks, follow the pattern used by TerminalBench2, TBLite, and YC-Bench:
1. **Create under** `environments/benchmarks/your-benchmark/`
2. **Set eval-only config**: `eval_handling=STOP_TRAIN`, `steps_per_eval=1`, `total_steps=1`
3. **Stub training methods**: `collect_trajectories()` returns `(None, [])`, `score()` returns `None`
4. **Implement** `rollout_and_score_eval(eval_item)` β the per-item agent loop + scoring
5. **Implement** `evaluate()` β orchestrates all runs, computes aggregate metrics
6. **Add streaming JSONL** for crash-safe result persistence
7. **Add cleanup**: `KeyboardInterrupt` handling, `cleanup_all_environments()`, `_tool_executor.shutdown()`
8. **Run with** `evaluate` subcommand
See `environments/benchmarks/yc_bench/yc_bench_env.py` for a clean, well-documented reference implementation.
## Configuration Reference
### HermesAgentEnvConfig Fields
| Field | Type | Default | Description |
|-------|------|---------|-------------|
| `enabled_toolsets` | `List[str]` | `None` (all) | Which hermes toolsets to enable |
| `disabled_toolsets` | `List[str]` | `None` | Toolsets to filter out |
| `distribution` | `str` | `None` | Probabilistic toolset distribution name |
| `max_agent_turns` | `int` | `30` | Max LLM calls per rollout |
| `agent_temperature` | `float` | `1.0` | Sampling temperature |
| `system_prompt` | `str` | `None` | System message for the agent |
| `terminal_backend` | `str` | `"local"` | `local`, `docker`, `modal`, `daytona`, `ssh`, `singularity` |
| `terminal_timeout` | `int` | `120` | Seconds per terminal command |
| `terminal_lifetime` | `int` | `3600` | Max sandbox lifetime |
| `dataset_name` | `str` | `None` | HuggingFace dataset identifier |
| `tool_pool_size` | `int` | `128` | Thread pool size for tool execution |
| `tool_call_parser` | `str` | `"hermes"` | Parser for Phase 2 raw output |
| `extra_body` | `Dict` | `None` | Extra params for OpenAI API (e.g., OpenRouter provider prefs) |
| `eval_handling` | `Enum` | `STOP_TRAIN` | `STOP_TRAIN`, `LIMIT_TRAIN`, `NONE` |
### YAML Configuration
Environments can be configured via YAML files passed with `--config`:
```yaml
env:
enabled_toolsets: ["terminal", "file"]
max_agent_turns: 60
max_token_length: 32000
agent_temperature: 0.8
terminal_backend: "modal"
terminal_timeout: 300
dataset_name: "NousResearch/terminal-bench-2"
tokenizer_name: "NousResearch/Hermes-3-Llama-3.1-8B"
use_wandb: true
wandb_name: "my-benchmark"
openai:
base_url: "https://openrouter.ai/api/v1"
model_name: "anthropic/claude-sonnet-4.6"
server_type: "openai"
health_check: false
```
YAML values override `config_init()` defaults. CLI arguments override YAML values:
```bash
python my_env.py evaluate \
--config my_config.yaml \
--openai.model_name anthropic/claude-opus-4.6 # overrides YAML
```
## Prerequisites
### For all environments
- Python >= 3.11
- `atroposlib`: `pip install git+https://github.com/NousResearch/atropos.git`
- An LLM API key (OpenRouter, OpenAI, or self-hosted VLLM/SGLang)
### For Modal-sandboxed benchmarks (TB2, TBLite)
- [Modal](https://modal.com) account and CLI: `pip install "hermes-agent[modal]"`
- `MODAL_TOKEN_ID` and `MODAL_TOKEN_SECRET` environment variables
### For YC-Bench
- `pip install "hermes-agent[yc-bench]"` (installs the yc-bench CLI + SQLAlchemy)
- No Modal needed β runs with local terminal backend
### For RL training
- `TINKER_API_KEY` β API key for the [Tinker](https://tinker.computer) training service
- `WANDB_API_KEY` β for Weights & Biases metrics tracking
- The `tinker-atropos` submodule (at `tinker-atropos/` in the repo)
See [RL Training](/user-guide/features/rl-training) for the agent-driven RL workflow.
## Directory Structure
```
environments/
βββ hermes_base_env.py # Abstract base class (HermesAgentBaseEnv)
βββ agent_loop.py # Multi-turn agent engine (HermesAgentLoop)
βββ tool_context.py # Per-rollout tool access for reward functions
βββ patches.py # Async-safety patches for Modal backend
β
βββ tool_call_parsers/ # Phase 2 client-side parsers
β βββ hermes_parser.py # Hermes/ChatML <tool_call> format
β βββ mistral_parser.py # Mistral [TOOL_CALLS] format
β βββ llama_parser.py # Llama 3 JSON tool calling
β βββ qwen_parser.py # Qwen format
β βββ deepseek_v3_parser.py # DeepSeek V3 format
β βββ ... # + kimi_k2, longcat, glm45/47, etc.
β
βββ terminal_test_env/ # Stack validation (inline tasks)
βββ hermes_swe_env/ # SWE-bench training environment
β
βββ benchmarks/ # Evaluation benchmarks
βββ terminalbench_2/ # 89 terminal tasks, Modal sandboxes
βββ tblite/ # 100 calibrated tasks (fast TB2 proxy)
βββ yc_bench/ # Long-horizon strategic benchmark
```
|